Measuring whether an AI system actually works, keeping it working, and evidencing that to someone who has to sign for it.
5 guides · updated 13 September 2026
01 · 12 min · 2026-03-16
Enterprise AI projects fail at an 80% rate — not because of bad models, but bad architecture. The 5 structural mistakes that account for nearly every failure.
02 · 12 min · 2026-09-09
The eval set is the durable asset and, in regulated work, the validation artefact. How to build one that is evidence rather than a spreadsheet.
03 · 12 min · 2026-08-28
LangSmith, Langfuse, Arize Phoenix, Braintrust and Datadog compared - and why the eval set you build matters more than the tool you buy.
04 · 14 min · 2026-09-09
What Part 11 actually requires of an LLM or RAG system - audit trails, e-signatures, and validating a component that is not deterministic.
05 · 13 min · 2026-09-10
Three different things are sold under this heading and buyers conflate them. Validation specialists, governance consultancies and builders compared, with criteria.
These write-ups come out of production systems rather than reading. If you are building in this area, here is how I work on it — and the benchmark results are the numbers behind the claims, including the ones that look bad.